Reduce Peak WAN inference VRAM usage - part II (Comfy-Org#10062)

* flux: math: Use _addcmul to avoid expensive VRAM intermediate

The rope process can be the VRAM peak and this intermediate
for the addition result before releasing the original can OOM.
addcmul_ it.

* wan: Delete the self attention before cross attention

This saves VRAM when the cross attention and FFN are in play as the
VRAM peak.

Author: rattus128

comfy/ldm/flux/math.py

@@ -37,7 +37,10 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
 
 def apply_rope1(x: Tensor, freqs_cis: Tensor):
     x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
-    x_out = freqs_cis[..., 0] * x_[..., 0] + freqs_cis[..., 1] * x_[..., 1]
+
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+
     return x_out.reshape(*x.shape).type_as(x)
 
 def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):

comfy/ldm/wan/model.py

@@ -237,6 +237,7 @@ def forward(
             freqs, transformer_options=transformer_options)
 
         x = torch.addcmul(x, y, repeat_e(e[2], x))
+        del y
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)